Mila

WebSwarm's innovative recursive delegation allows agents to not only search but also adaptively collaborate, leading to superior performance in complex web search tasks.

The probability of a model trajectory entering unsafe regions can be exponentially small, but the geometry of the unsafe set critically influences how quickly training stabilizes.

Adversarial training on just 66 carefully chosen statements can achieve near-zero attack success rates, revolutionizing safety alignment in language models.

Fast transductive rates in semi-supervised learning can be achieved with fewer labels than previously thought, thanks to the power of data augmentation.

Retaining past knowledge can actually impede real-time adaptation in dynamic environments, leading to a new framework for optimizing continual learning.

LLMs possess an internal estimate of their remaining output length, revealing a surprising layer of planning in their generation process.

Even advanced LLMs struggle to prevent privacy breaches in multi-user settings, exposing critical data spillage risks that current benchmarks overlook.

Trust-region optimization can dramatically enhance the training of neural quantum states, achieving stability and speed at unprecedented scales.

Gemma 4's unified architecture and reasoning mode enable it to outperform larger models in human-rated tasks while maintaining high efficiency.

Gradual transitions in training objectives can significantly enhance model performance during adaptation, preserving valuable learned features.

Spectral Envelope Theory unlocks a new level of fidelity in generating synthetic relational time series, outperforming traditional methods in capturing complex temporal dynamics.

ArBG achieves a remarkable 60% reduction in zero-shot energy error for peptide systems, challenging the dominance of flow-based sampling methods.

Combining hard architectural sparsity with soft regularization can significantly enhance the interpretability of vision models without sacrificing performance.

Reusable fixing transformations can achieve a 94.3% compilable transformation rate, revolutionizing how we handle breaking API changes across multiple projects.

Iteratively training on a self-selected dataset can dramatically enhance vision-language model performance without the need for extra data or pre-training.

Allocating more capacity to earlier layers in language models can significantly enhance performance, challenging the long-held uniform layer design paradigm.

FlowMaps can revolutionize robotic navigation by accurately predicting object movements based on human interaction patterns, outperforming existing methods in real-world scenarios.

N-version programming with coding agents not only mirrors historical failures but also shows a dramatic reduction in errors through diversity, challenging assumptions about AI reliability.

Discriminator-Guided RL achieves a remarkable reduction in FID from 9.38 to 2.62, showcasing a new way to align model outputs with real data without human preferences.

Scaling self-play training directly from pixels leads to competitive autonomous driving performance without human trajectory supervision.

LSTM decoders can boost attention alignment with human fixation patterns by up to 50 percentage points, but at the cost of spatial precision and task differentiation.

Negative token filtering enables stable single-rollout training, outperforming traditional group-based methods on agentic tasks while maintaining efficiency.

MambAdapter achieves superior performance in audio and speech tasks while drastically cutting down on computational resources.

STORM transforms lexical query expansion by turning retrieval rewards into actionable token-level signals, enabling efficient and effective query rewriting that rivals larger models.

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

Freezing parameter groups based on Fisher structural drift can enhance fine-tuning efficiency and boost performance in downstream tasks.

No single TTS model excels across all languages, exposing the limitations of current multilingual approaches in low-resource settings.

SHAPO redefines safe exploration by biasing policy updates toward caution in uncertain regions, leading to significant improvements in safety and performance.

Systematic gaps in AI evaluation reporting are exposed, revealing inconsistencies that hinder reliable comparisons across thousands of models and benchmarks.

Prefix failure in on-policy distillation can be effectively mitigated by correcting problematic prefixes, leading to significant improvements in reasoning coverage and accuracy.

Active exploration can dramatically enhance adults' ability to reason about complex causal relationships, but even with this advantage, they still struggle compared to simpler tasks.

Code-switched speech can exploit safety weaknesses in LALMs, achieving jailbreak success rates that challenge current safety protocols.

Counterfactuals, often seen as safe tools for model explanation, can inadvertently expose sensitive training data through membership inference attacks.

AuditFlow achieves over 82% accuracy in structured financial audits by leveraging a unique symbolic environment that outperforms traditional methods by nearly 15 points.

MobEvolve outperforms traditional methods by evolving its logic through targeted updates, achieving unprecedented fidelity and interpretability in human mobility generation.

Bridging the scientific knowledge gap for hundreds of millions, AfriScience-MT pioneers document-level scientific machine translation for six African languages.

How you represent a plan matters more than which LLM you use when building robust web agents.

Language specialization in multilingual MoEs happens mostly in the final layers, suggesting a surprisingly simple recipe for parameter-efficient adaptation.

Human-generated citation lists, long considered the gold standard for evaluating literature search, are surprisingly unreliable, with LLMs judging them relevant only ~50% of the time.

Offline policy optimization with a world model allows for affective music recommendation that improves user valence and arousal, even when ethical constraints preclude online experimentation.

Uniformly quantizing the entire diffusion action head of VLAs to W4A4 is not only possible, but can match or exceed FP16 performance, defying conventional wisdom and slashing memory footprint by 71%.

Exploiting approximate label symmetries can significantly improve data scaling in machine learning, even when those symmetries aren't perfect.

Frustrated by biased algorithms? Now, users can band together to nudge AI systems toward fairness, even without the platform's permission.

Forcing VLMs to "think" visually with panoramic renderings, rather than relying on language alone, unlocks surprisingly robust spatial reasoning.

Causal features aren't just for understanding data; they can also align long-term incentives in strategic classification, offering a path to robustness and fairness.

As AI agents scale and interact, the Foundation Protocol offers a coordination layer that prioritizes accountability and governance, ensuring that the future of human-AI collaboration remains open and governable.

DINOv2's representation space is so statistically well-behaved that you can train a vanilla diffusion transformer on it and beat specialized architectures with fewer parameters.

Forget expensive downstream evaluations: token-level statistics from expert-written solutions can reliably forecast LLM performance with 10,000x less compute.

Dye-filled boron nitride nanotubes aren't the optically-bright J-aggregates we hoped for, but instead form structurally heterogeneous confined ensembles with suppressed radiative rates.

LLM-powered query reformulation, a hot topic in IR, often fails to translate gains from lexical to neural retrieval, and bigger models don't always help.

LLMs struggle with structured 2D tasks when inputs are serialized into 1D, revealing a surprising performance gap compared to vision-augmented models that directly process the 2D layout.

AI is poised to revolutionize protein dynamics research, but key challenges remain in ensuring scalability, thermodynamic consistency, and kinetic fidelity.

CroSearch-R1 reveals that integrating cross-lingual knowledge through a dynamic retrieval strategy can substantially enhance the performance of Retrieval-Augmented Generation systems.

ManifoldRank reveals that treating fairness as a taxation cost can significantly enhance the effectiveness of online fair re-ranking algorithms.

LLMs re-rank documents better when you learn to route each query to the specific attention heads that matter, instead of relying on static subsets or everything at once.

Sampling plausible configurations of digital twins can reveal multiple valid parameterizations, enhancing model adaptation in complex natural systems.

Out-of-domain self-supervised pretraining on brain MRIs beats in-domain supervised learning when generalizing to real-world clinical data.

Stop overpaying for LLM serving: intelligently routing requests to specialized pools based on token budget slashes GPU costs by up to 42% and dramatically improves reliability.

Multi-modal alignment in symbolic regression models like SNIP doesn't actually improve during optimization, suggesting current approaches are too coarse to effectively guide symbolic search.

Merging experts in MoE LLMs can actually *improve* performance compared to pruning, offering a new path to compression that preserves capabilities.

Forget hand-crafted prompts: RL can automatically unearth 36 new failure modes in VLMs that humans miss, revealing surprising blind spots in counting, spatial reasoning, and viewpoint understanding.

MoEs don't always need learned routers: routing information can be embedded directly in the hidden state.

Sparse autoencoders' failure to generalize compositionally isn't due to amortized inference, but because they learn lousy dictionaries in the first place.

Mimicking human cognition, FLAIR lets dialogue models "think while listening," boosting performance without adding latency.

Forget exotic attention mechanisms – MobileLLM-Flash achieves up to 1.8x faster LLM prefill on mobile CPUs by smartly pruning and adapting existing architectures for on-device use.

Ditch the cross-world counterfactuals: Sequential Transport offers a DAG-aware, optimal transport approach to causal mediation analysis, providing deterministic counterfactual mediators and fine-grained attribution.

Tactile sensing can be efficiently injected into vision-language-action models via feature-wise linear modulation, boosting robot manipulation performance without the computational overhead of large-scale pretraining.

Forget contrastive learning: LLM2Vec-Gen learns text embeddings by representing the *response* an LLM would generate, unlocking safety and reasoning abilities for embedding tasks.

Democratized LLM pre-training is now a reality: Covenant-72B proves you can train a competitive 72B model with untrusted peers over the internet, opening the door to broader participation and reduced costs.

Diagonal SSMs, despite their empirical success, provably fail to track states of non-Abelian groups, revealing fundamental limitations in their expressive power.

One in four initial posts on a major cybercrime forum contain explicit crime-related content, revealing a surprisingly high baseline of open criminal activity.

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

Forget full fine-tuning: this dynamic routing strategy lets you adapt dense retrieval to new domains while using just 2% of the parameters.

Quantum computers could compromise nuclear power plants' safety systems and create unsolvable forensic paradoxes, with current defenses showing alarming vulnerability.

A global consensus on AI safety risks and capabilities has emerged from a panel of 100+ independent experts, representing a landmark effort in international collaboration.

Achieve state-of-the-art dynamic graph anomaly detection with limited labels by learning a robust decision boundary around normal data, outperforming methods that overfit to scarce anomalies.

Attention-based re-ranking gets a boost: ReAttn's post-hoc re-weighting tames over-concentration and lexical bias, leading to more accurate and interpretable results without extra training.

LLMs struggle to balance rational financial decisions with mimicking noisy user behavior, often overfitting to short-term market trends instead of aligning with long-term investment goals.

Coreference benchmarks may be overstating language models' NLU abilities, as even small changes to evaluation contexts reveal a failure to generalize.

Forget fixed teams: this new reinforcement learning framework lets agents spawn new teammates on the fly, unlocking dynamic strategies previously impossible.

Boost macrocycle generation rates from 1% to 99% by guiding diffusion models with persistent homology, opening new avenues for drug discovery.

Cybercriminals are actively exploring AI's potential for both enhancing existing attacks and creating novel illicit tools, but harbor significant doubts about its real-world effectiveness and impact on their operations.

Freehand sketches can now drive photorealistic image generation, even without paired training data, thanks to a novel loss that prioritizes semantic understanding over pixel-perfect alignment.

Current multimodal retrieval systems fall flat when faced with realistic visual streams where context is distributed across time, motivating a new agentic paradigm for context-aware image retrieval.

Mismatched standard deviations in multi-objective RL advantage estimation can completely break constrained learning, but a simple scalarization fixes it.

This study establishes SSL as a promising paradigm for ECG analysis, particularly in settings with limited annotated data, enhancing accessibility, generalizability, and fairness in AI-driven cardiac diagnostics across diverse clinical environments and questions.

Even when trained on suboptimal data, a Bayesian in-context RL agent can achieve near-optimal decisions on unseen tasks by fusing a learned Q-value prior with in-context information and employing an upper-confidence bound for exploration.

GPT-5's real-time router learns to route queries to specialized models, making it faster and more useful than its predecessors.

This work integrates small-molecule high-throughput screening with a deep-learning-based virtual screening approach to uncover new antibacterial compounds, illustrating a 90-fold improved hit rate over the high-throughput screening experiment used for training.

By recursively aggregating reasoning chains, even smaller LLMs can now achieve performance competitive with much larger models, challenging the assumption that scale is the only path to improved reasoning.

Dramatically improve protein language models by simply post-training them to align with protein graphs, yielding a 59% increase in contact prediction accuracy.

Command A shows how to build an enterprise-grade LLM that balances performance, efficiency, and multilingual capabilities using decentralized training and model merging.

Forget retraining from scratch: port fine-tuning updates between LLM versions and get up to 47% performance boost on tasks like instruction following, even surpassing fully fine-tuned models.

Self-supervised learning beats supervised learning for ECG interpretation when labeled data is scarce, unlocking more robust and generalizable AI-driven cardiac diagnostics.