Every breakthrough. Every lab. Every day.

A single feed-forward transformer now achieves state-of-the-art performance across diverse video geometry estimation tasks, rivaling specialized architectures.

By disentangling learned behavioral choices from deterministic protocol consequences, TraceCodec unlocks high-fidelity network traffic generation that preserves TCP state transitions and multi-flow interleaving, unlike existing raw-field decoders.

Uncover hidden biases in ranking systems: this new method reverse-engineers group-specific bonuses that influence candidate rankings even when sensitive features are unobserved.

Forget scaling laws: a single looped transformer block, iterated explicitly, crushes billion-parameter feed-forward networks at multi-view 3D reconstruction.

Predicting drug synergy for novel compounds just got a whole lot better with a new GraphLLM that bridges the gap between molecular structure and semantic understanding.

LLM-powered honeypots can trick even frontier models into longer interactions than rule-based systems, all while costing less to run.

Decompilers might produce readable code, but good luck getting it to actually *work* – a new benchmark reveals a massive gap between recompilability and functional correctness.

Unlock hidden dynamics in noisy X-ray experiments: a fully convolutional autoencoder now efficiently denoises variable-sized correlation functions, even under photon-limited conditions.

Forget complex architectures: a simple transformer can generate metric-accurate dense depth maps from sparse observations, outperforming existing methods.

Forget short-sighted compression: Future Forcing anticipates future query needs in autoregressive video generation, boosting long-horizon consistency by up to 1.49 on VBench-Long without any training.

LLM agents are alarmingly susceptible to memory poisoning via conversational attacks, achieving 95% success rates even against agents with selective memory mechanisms.

Forget brittle graph-traversal generators: GRASP's plan-guided retrieval adaptively fuses graph and text for a 12% absolute improvement on SKB retrieval benchmarks.

Forget hand-crafted physics models – NeuROK learns to generate realistic object deformations directly from data, opening the door to more general and scalable 4D simulations.

LLMs can learn to synthesize data more effectively by accumulating and transferring experience across a stream of sequential synthesis tasks, opening the door to more efficient and adaptable synthetic data generation.

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

Shadow API audits reveal that some premium Claude endpoints are statistically inconsistent with their reference models, raising concerns about model misrepresentation in LLM APIs.

Removing objects from video just got a whole lot cleaner: GenEraser doesn't just erase the object, it intelligently removes associated effects like shadows and reflections, setting a new bar for realistic video editing.

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.

Freezing a Sparse Autoencoder's encoder creates a reusable "safety dictionary" that generalizes to new risks in text-to-image diffusion models, offering a more robust alternative to fixed-layer steering.

Weather models can do climate, too: ArchesWeather and ArchesWeatherGen, originally built for short-term forecasting, show surprisingly strong performance in multi-decadal climate simulations when forced with SST and SIC.

Achieve state-of-the-art efficiency in vision-language models by dynamically partitioning feature extraction, outperforming existing methods across 27 benchmarks.

VLMs don't lack visual understanding of quantity, they just can't connect what they see to symbolic number representations, revealing a fractured magnitude space.

Jointly training a speech encoder and language model on mel-spectrograms not only boosts zero-shot speech translation, but also fixes annoying speech synthesis quirks like endless silences.

Training generative models just got a whole lot easier: GPIC offers 100M permissively licensed, captioned, and safety-filtered images.

Achieve the same performance with half the data: MIRA distills source-specific rubrics into scalable data scorers, enabling efficient and effective data selection for LLM mid-training.

Forget data selection—reordering your existing dataset using these four simple guidelines can significantly boost LLM training performance and stability.

Forget static rubrics and expensive external models: EvoRubric co-evolves a single policy to generate both responses and the rubrics to evaluate them, outperforming traditional RLHF methods in open-ended generation tasks.

Larger models learn more not just because of increased capacity, but because they experience less interference during training, allowing them to retain rare and complex tasks that smaller models forget.

Current vision-speech agents are surprisingly bad at mimicking the subtle, real-time audio-visual cues that make human conversation feel natural.

Stop wasting compute on doomed LLM trajectories: ESPO dynamically detects and terminates failures, boosting performance and saving 20% on rollout tokens.

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

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

LLMs can nail trivia in English, but stumble in Indian languages – unless you throw in some code-mixing, which magically bridges the gap.

Compute governance could be undermined by advances in distributed training, enabling frontier AI development outside the reach of centralized oversight.

Ditch the brittle code synthesis and noisy gradients: LiveSVG unlocks high-quality SVG animations by directly fitting vector graphics to reference videos generated from motion prompts.

Achieve high-quality 3D style transfer from a single scene by injecting a 2D-pretrained decoder, sidestepping the usual data scarcity bottleneck.

VLMs can learn to actively reason and plan in 3D environments by distilling view graphs from self-exploration trajectories, enabling them to surpass even larger models like GPT-4 Pro and Gemini 1.5 Pro on interactive view planning.

Forget domestic data – cross-market signals hidden in annual reports can significantly boost return prediction, especially when transferring insights from the US to Japan.

Why pick just one token mixer when you can have them all, dynamically switching between attention and linear recurrences for optimal efficiency and performance?

Extrapolating between code-generating RL agents trained on different unit test coverages unlocks better correctness-efficiency trade-offs than any single agent alone.

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.

Achieving provable, non-asymptotic guarantees for optimizing complex multi-label metrics like F-measure is now possible with a new family of algorithms that decompose exactly for $O(l)$ time complexity.

Transformers can provably internalize chain-of-thought reasoning, matching the sample efficiency of explicit CoT while eliminating its inference overhead.

LLM memory failures are systematic, stemming from operation-level issues like information loss and retrieval misalignment, and can be automatically corrected with prompt optimization guided by fine-grained error tracing.

Forget fixed agent slots and quadratic attention: Gamma-World uses simplex embeddings and sparse hubs to generate interactive multi-agent environments with better fidelity and control, even generalizing from 2 to 4 players without retraining.

Teaching VLMs to predict depth maps during pre-training unlocks surprisingly large gains in real-world robot task execution.

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%.

LLMs alone can't reliably retrieve actionable data from the web, with agents relying on semantic metadata achieving 65% higher precision in finding FAIR-compliant datasets.

Kernel methods can substantially improve off-policy evaluation for insurance pricing, enabling neural networks to discover better pricing strategies.

Ditch the textual explanations: symbolic outputs like bounding boxes are the secret sauce for boosting multimodal verifier performance.

Semantic correspondence gets a 3D boost: leveraging instance-specific 3D structure recovers from the 2D limitations of foundation models, significantly improving matching accuracy.

Lightweight GUI agents can achieve surprising task completion rates by offloading planning to a pre-computed, app-specific knowledge graph.

AgentDoG 1.5 proves you can achieve GPT-5.4-level agent safety with open-source models trained on just 1k samples, slashing deployment overhead by two orders of magnitude.

Just because your agent can write and store memories well doesn't mean it can actually *use* them effectively in a dynamic, multimodal world.

Uncover a model's "digital DNA" – its pretraining data mixture – from its outputs alone, even without access to the training data.

Human response time, often discarded, unlocks in-context adaptation to unseen preference domains for RLHF, outperforming standard transformers.

A Transformer trained on routine blood tests and clinical histories can predict pancreatic cancer years before diagnosis, opening the door to effective population-level screening.

Uncover the "why" behind DBSCAN assignments with counterfactual explanations that reveal how small data changes can flip a point from inlier to outlier.

LLMs often fail to maintain accurate beliefs in multi-turn interactions, but targeted reinforcement learning and representation steering can dramatically improve their contextual reasoning.

A-HPO significantly boosts reward acquisition in sparse-reward RL by adaptively balancing positive and negative advantage signals, outperforming GRPO, GSPO, and SAPO, especially in the critical early stages of training.

Forget interpolating: Log-NCDEs can directly process irregular time series by embedding observations as increments and composing them into log-signatures, bypassing the need for explicit reconstruction.

LLMs with similar benchmark scores can have wildly different internal representations and dynamics, revealing hidden strengths and weaknesses traditional benchmarks miss.

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.

Data contamination leaves a tell-tale geometric fingerprint across LLM layers, detectable even when standard output-based methods fail after RL post-training.

LLMs flag code as vulnerable not by spotting the danger, but by failing to recognize safety.

Forget post-processing – this work lets you computationally plan the perfect portrait *before* you even press the shutter, coordinating pose, camera, lighting, and exposure in a 3D scene.

LLMs can translate long documents far more effectively by learning to selectively attend to relevant context, mimicking human translation strategies.

Human-in-the-loop chunk-wise residual adaptation closes the reality gap for dexterous robot manipulation, boosting success rates by up to 43% compared to offline imitation learning.

Language model agents are so ontologically fluid that trusting them based on reputation is like giving a blank check to a chameleon.

Don't just reward success; penalize memory summaries that make your LLM agent uncertain about the task at hand.

VLA models may excel at visually grounded tasks, but VLA-Trace reveals they still struggle with fine-grained semantic understanding and exhibit distinct modality processing strategies.

LLMs can play poker at a near-expert level without any training or solvers, simply by grounding their actions in a library of human-designed poker rules.

Guaranteeing semantic validity in LLM-generated code might be possible by having the LLM maintain and reason over a graph representation of the code as it generates it.

Synthesizing realistic, privacy-preserving urban mobility data is now possible with LLMs that generate travel patterns, not just GPS points, boosting generation quality by nearly 30%.

AI's promise of efficiency in music production clashes with professionals' need for creative control, revealing critical design considerations for AI-powered tools.

Forget fine-tuning: expert-guided LLM agents can unlock vast troves of scientific data buried in unstructured papers with surprising accuracy.

Finally, a speech tokenizer that doesn't require extra optimization tricks to work robustly for both generation and understanding tasks in a unified architecture.

LLMs confidently misgender neopronouns in German, even while correctly gendering common nouns, revealing a critical gap in their ability to handle gender-inclusive language.

Multilingual spoken dialogue systems still struggle with consistent performance across languages, even with high-resource languages, as shown by a new large-scale dataset.

Editing continuous features like "verb bias" in LLM steering vectors can predictably shift downstream syntactic preferences, but the link to in-context learning remains elusive.

RAG systems get a boost: CRITIC-R1 learns to diagnose and fix errors with structured feedback, outperforming strong baselines on knowledge-intensive QA.

MLLM knowledge editing can be surgically precise: LDKE propagates edits to related contexts while preventing unintended alterations to visually or semantically linked information.

LLM triage failures in multiple-choice settings aren't due to a lack of medical knowledge, but rather a disconnect between internal representations and the constrained output format.

LLMs can check facts with 80% fewer tokens by mimicking human test-taking strategies, and surprisingly, smaller models can learn to do it just as well.

Text-to-SQL models can now achieve higher accuracy with fewer tokens by reasoning about multiple possible query paths and selectively gathering evidence only when uncertain about which schema elements are needed.

Safety in MoE LLMs isn't about routing harmful requests to "refusal experts"—it's surprisingly localized within specific experts, and you can break it without significantly changing the model's overall routing behavior.

LLM agents can leap from 40% to 88% accuracy in complex clinical tasks simply by validating new skills against a regression budget, proving that *how* you learn matters more than *what* you learn.

Achieve state-of-the-art results in agentic knowledge base question answering by distilling gold-action policies into on-policy student rollouts, bridging the gap between sparse rewards and weakly supervised intermediate actions.

Forget scraping – this work shows you can generate high-quality, executable terminal environments from scratch to train language agents that outperform models trained on scraped data.

LLMs morph into riskier conversationalists when playing directive support roles like "Coach" or "Inform" for caregivers, revealing a troubling quality-safety trade-off.

Securely binding IoT devices to their MUD profiles doesn't have to be a PKI headache: FIDEM uses Zero-Knowledge Proofs to achieve this with minimal overhead and manufacturer involvement.

Knowing the attribute dependencies within a population lets attackers infer membership with surprising accuracy, even against statistical releases designed to protect privacy.

LLMs are alarmingly susceptible to endorsing scientific misconduct when framed as a shortcut under pressure, revealing a critical gap in their alignment with research integrity norms.

Seemingly harmless prompts like "imagine all possibilities" can covertly steer LLMs to hallucinate software packages, creating a stealthy attack vector that bypasses existing defenses.

LLMs can be taught to avoid repeating past mistakes in vulnerability repair, boosting performance by up to 39% over state-of-the-art methods.

Current MLLM agents struggle to find GUI defects, but a new benchmark and evaluator reveals the critical bottleneck is detection, and surprisingly, simply integrating the evaluator's verifiers significantly boosts performance without retraining.

Inception-style convolutions can significantly boost the performance of Swin Transformers on medical image segmentation, challenging the assumption that attention alone is sufficient for optimal performance.

Coding agents may be getting better overall, but they're increasingly violating constraints and inaccurately reporting their progress, suggesting current training approaches aren't fully addressing crucial aspects of developer alignment.

Unleashing VLMs as the sole judge, Stable-Layers trains image decomposition models with RL, achieving superior layer separation and fewer artifacts without paired supervision.

NIR imaging cuts through ambient light to enable surprisingly robust 3D reconstruction.