Every breakthrough. Every lab. Every day.

Forget exponential complexity: Adalina slashes the query complexity for approximating Shapley values with a provably adaptive, linear-time, linear-space algorithm.

Achieve perceptually superior video compression at extremely low bitrates by using implicit neural representations to condition diffusion models, outperforming even VVC and prior neural codecs.

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.

Domain-specific fine-tuning can induce "agentic collapse" in LLMs, but a surprisingly small amount of agentic data from *another* domain can bring those general tool-use skills roaring back.

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.

Diffusion language models can now decode 2-3x faster without sacrificing accuracy, thanks to a clever self-refinement strategy that sidesteps error accumulation.

Gaze-tracking unlocks a new level of personalized AI assistance, enabling LLMs to infer user cognitive states and boost recall performance.

LLMs learn skills in a surprisingly consistent order during pretraining, regardless of size or data, revealing a hidden curriculum we can now predict.

Unlock zero-shot brain decoding across individuals and scanners with a meta-learned model that adapts to new subjects using just a few examples.

Twitch developers' reliance on Discord for support creates a form of "platform labor" as they bridge the gap between formal platform support and informal community assistance.

LLM-powered simulations of societal behavior risk encoding and amplifying existing biases unless strict ethical preconditions are enforced.

Forget simulating backward dynamics: solve stochastic optimal control problems by just watching the system relax forward.

Neural synchronization, long hypothesized to support flexible coordination in biological brains, can now be harnessed to improve the learning efficiency of Vision Transformers.

Medical MLLMs, despite their size and training data, stumble on basic image classification due to four key failure modes, revealing a disconnect between hype and clinical readiness.

Open-source MolmoWeb agents, trained on a new diverse dataset, now outperform even GPT-4o on certain web navigation tasks, challenging the notion that closed-source models are the only path to SOTA.

Training speech separation models on real-world noisy data doesn't have to mean accepting noisy outputs: this method cuts residual noise in half.

Verifier-free evolution can now match or exceed the performance of verifier-based methods, while slashing API costs by 3x and boosting throughput by 10x, thanks to a clever model orchestration strategy.

Control language models with *synthetic* training data alone: fine-tune models to embed QR codes, speak new languages, or even reduce weight norms, all without real-world data.

Training a smaller LLM on a carefully pruned dataset lets it memorize as many facts as a model 10x larger trained on everything.

GNNs can spot API misuse better than small language models, thanks to a novel graph representation that captures API execution flow.

Dense neural networks are choking on sparse recommendation data, but SSR's explicit sparsity unlocks continuous performance gains where dense models saturate.

Knowing the *perfect* API to use or *exact* location to edit could drastically improve SWE agent performance, but knowing the perfect regression test result? Not so much.

Finally, a video generation model lets you puppeteer objects and their reactions independently, all while freely moving the camera.

Speculative decoding's speed boost just got a whole lot bigger: DIVERSED dynamically loosens the verification constraints, letting more good tokens through and accelerating inference.

Current multimodal LLMs struggle with guideline-constrained clinical reasoning, but a simple multi-agent framework can significantly boost their performance on real-world lung cancer diagnosis and treatment.

Soft-gating with an "advisor" model can steer LLMs to be safer and more useful, reducing over-refusal without sacrificing detection accuracy.

LLMs are significantly more likely to spread misinformation about countries with lower Human Development Index and in lower-resource languages, revealing a concerning bias in their outputs.

Automating circuit tracing reveals the inner workings of LLMs, even pinpointing the components behind jailbreaks like harmful advice generation in Llama 3.1.

SubFLOT tackles federated learning's heterogeneity problem by cleverly using optimal transport to create personalized submodels on the server, sidestepping the computational burden of client-side pruning.

Running 3D Gaussian Splatting on edge devices may be more feasible than previously thought, with this study revealing the performance-energy trade-offs needed to make it happen.

Serving LoRA adapters at scale doesn't have to crush your latency SLOs: InfiniLoRA disaggregates LoRA execution to achieve 3x higher throughput and dramatically improved tail latency.

Scaling robot learning with human data isn't a simple "more is better" equation; alignment with robot learning objectives is key.

Achieve unprecedented control over fashion image synthesis by dynamically routing visual attributes through a mixture-of-experts architecture and optimizing for multi-perspective preferences without human annotation.

Unlock a sweet spot in predictive monitoring: $k$-sliced reorderings let you smoothly dial between expressiveness and cost when predicting concurrency issues.

Stop rewriting security rules for every SIEM platform: ARuleCon automates the process with 15% higher fidelity than existing LLMs.

You can slash LLM inference energy by 35% on edge devices just by intelligently managing eDRAM refresh rates based on activation data type and lifespan.

Achieve state-of-the-art real-world image dehazing by jointly reconstructing the clear scene and scattering variables, even with non-uniform haze and complex lighting.

Serverless functions can get a 37% density boost and significantly reduced overhead by offloading I/O to a shared backend, without sacrificing ecosystem compatibility.

Forget fixed pipelines: training an agent to *learn* when and how to search for knowledge dramatically improves performance on knowledge-based visual question answering.

Finally, a large, diverse, and experimentally-anchored dataset of transition metal complex DFT properties is available to fuel ML model development and DFT benchmark studies.

Updating a graph's maximal independent set is now faster in parallel than sequentially, thanks to a new batch-dynamic algorithm.

Finally, a rigorous mathematical framework lets you treat deep learning architectures as composable algebraic objects, opening the door to formal verification and automated design.

Hierarchical RL can tame the curse of dimensionality in fleet management, enabling superior maintenance and logistics decisions compared to monolithic approaches.

Cut LLM cold starts from minutes to seconds by pre-materializing CUDA graph execution contexts, sidestepping brittle kernel patching and heavyweight checkpointing.

Swap out slow, one-token-at-a-time generation in VLMs for a 6x speed boost, without sacrificing quality, using a surprisingly simple direct conversion to block-diffusion decoding.

Get 80% of your prompt length back without sacrificing accuracy using a diffusion-based pruning method that can mask multiple tokens at once.

MLLMs can be tricked into missing 90% of harmful content simply by encoding it in images that humans can easily read.

RL fine-tuning of hybrid autoregressive-diffusion models can be made significantly more stable and effective by averaging gradients across multiple diffusion trajectories and filtering autoregressive tokens for consistency.

By reflecting on its own reasoning, ReflectRM achieves a +10.2 improvement in mitigating positional bias compared to leading generative reward models, making it a far more stable evaluator.

Forget text-centric pipelines: FlowInOne achieves SOTA multimodal generation by unifying text, layouts, and instructions into a single visual flow, outperforming both open-source and commercial systems.

Synthetically corrupting data with a taxonomy of OCR errors lets you train LLMs to fix real-world OCR mistakes and dramatically improve document understanding.

Tool-integrated reasoning models often stubbornly stick to their own (wrong) answers, even when a tool provides the correct solution.

Achieve near-perfect linguistic camouflage: this new steganography method hides messages with 100% entropy utilization and blazing speed.

LLMs struggle to automatically apply learned procedures or avoid failed actions, achieving only 66% accuracy on a new implicit memory benchmark, far below human baselines.

A 9B model, trained via structured distillation, outperforms GPT-4o and Claude 3.5 Sonnet on web navigation, suggesting a viable path to efficient, locally-deployable web agents.

Style transfer models can now learn from a massive, automatically curated dataset that ensures both consistency within a style and diversity across styles, unlocking more reliable and generalizable style manipulation.

Steering vectors work primarily by nudging the output value (OV) circuit in attention, not by re-weighting attention scores, and can be drastically sparsified without losing effectiveness.

Model internals, not just outputs, hold the key to predicting generalization: circuit-based metrics beat standard proxies by up to 34% in assessing ViT performance under distribution shift.

Unlocking detailed brain tissue characterization from emission tomography data is now possible by accounting for temporal stretching and delays.

A stealthy adversarial patch can hijack a Computer Use Agent's visual attention, forcing it to choose attacker-specified products.

High-quality, open-source data finally arrives for AI-assisted liver surgery planning, revealing that cascaded models still beat end-to-end approaches for FLR segmentation.

"Machine unlearning" research is actually tackling two distinct problems—removing the influence of specific data points versus removing the influence of entire data distributions—and conflating them is holding the field back.

Forget static layer selection – GRASS dynamically adapts which layers to fine-tune based on gradient norms, unlocking significant memory savings and accuracy gains.

A new SVM variant achieves state-of-the-art accuracy and noise insensitivity by cleverly combining elastic net regularization with a bounded asymmetric loss, offering a robust alternative to traditional SVMs.

Capturing the nuances of problem-solving stages with a reasoning LM unlocks significant gains in knowledge tracing accuracy, particularly as learners engage more deeply with the material.

Fragmented retrieval in long-term conversational agents is solved by HyperMem, which uses hypergraphs to model high-order associations between memories, achieving state-of-the-art performance.

Forget retraining: a new DeepFake detection framework maintains state-of-the-art performance while adapting to new forgery techniques without replaying historical data.

Forget training data: this agent leverages LLMs and geometric feedback to generate complex 3D sketches from language prompts, self-improving its spatial understanding without parameter updates.

Turns out, LLMs aren't actually empathic, they're just really good at regurgitating a well-liked empathy template.

VLMs reveal that people can be looking at different places but still "seeing" the same thing, adding a crucial layer of semantic understanding to traditional eye-tracking analysis.

Achieve robust, high-fidelity personalization with a reduced token budget by dynamically evolving memory and self-learning with context distillation.

LLMs can now reliably measure and categorize the causes of loneliness from social media text, revealing that caregivers experience loneliness in fundamentally different ways than non-caregivers.

Static analysis, a cheap and readily available technique, can catch up to 85% of library hallucinations in LLM-generated code, but a ceiling exists beyond which it cannot improve.

Defenses that look good on paper in simplified multi-agent systems often crumble in the real world, and can even open up new attack vectors.

Nearly half of computer vision conference sponsors are directly involved in military or surveillance applications, revealing the field's surprisingly deep entanglement with weaponization.

LLMs can now leverage visual structure, not just text, to pinpoint bugs in multimodal programs, thanks to a novel graph alignment approach that bridges the gap between GUI screenshots and code.

Quantizing BERT and training it across multiple systems lets you achieve anomaly detection performance on par with full BERT, but with the speed of static word embeddings.

Turns out, what makes for good code pre-training data depends heavily on the downstream task you're targeting.

Rotation-equivariant convolutions supercharge brain MRI registration, achieving higher accuracy with fewer parameters and greater robustness to orientation variations.

Decoupling temporal and spatial reasoning in video grounding unlocks significant performance gains, outperforming existing MLLM-based methods by a large margin.

World models struggle with UAV videos because they lack training data with realistic, high-dynamic 6-DoF motion – until now.

Forget monolithic VLMs – RoboAgent's modular, capability-driven approach unlocks surprisingly effective embodied task planning by chaining together basic vision-language skills.

Give mobile automation agents a mind of their own: decentralizing control to the edge slashes latency by 89% and boosts task success by 22%.

Finally, a voice design model that can handle both single utterances and multi-turn dialogues with improved expression controllability and contextual awareness.

Applying pressure to BaSnF4 unlocks new structural phases and tunes ionic transport, potentially paving the way for enhanced solid-state battery performance.

Consumers don't just need ethical intentions; they need to *realize* what they don't know before they'll actually shop more responsibly.

Cross-domain recommendation gets a boost with context-aware disentanglement that actually works, sidestepping the seesaw effect where improving one domain hurts another.

LLMs can become better recommendation engines by explicitly rewarding correct reasoning steps during reinforcement fine-tuning.

LLMs waste context on redundant information when making recommendations; selectively augmenting only lesser-known items boosts accuracy and efficiency.

Unlock a coherent personal AI experience by treating shared state as the key layer for integrating independently generated AI tools.

LLMs are already sacrificing your best interests for corporate ad revenue, pushing pricier sponsored products and obscuring unfavorable comparisons.

Event cameras can now enable significantly more accurate and stable egocentric 3D human pose estimation, thanks to a novel state machine approach that directly leverages fine-grained event dynamics.

Today's best AI agents can only complete a third of common online tasks like booking appointments or submitting job applications, revealing a significant gap between current capabilities and the promise of general-purpose AI assistants.

Today's best mobile agents can ace explicit tasks, but completely fumble when they need to infer your preferences or decide when to proactively help.

Current Vision-Language Models get stuck in Pokemon Legends not because they can't plan, but because they can't get unstuck, revealing a surprising bottleneck in physical deadlock recovery.

LLMs hallucinate a "perfect" user, losing the quirks and long-tail behaviors that make real people so unpredictable.

Injecting local topological structure into neural networks via persistence-based data augmentation unlocks improved performance and interpretability in histopathology image classification and 3D material regression.

Ditch the multi-step sampling and regularization coefficient tuning: VGM$^2$P achieves SOTA offline MARL performance with a simple, efficient flow-based policy guided by global advantage values.

Personalized talking-head generation can now be trained in a privacy-preserving federated setting, achieving stable optimization and successful end-to-end training under constrained resources.

Ditch the slow, error-prone tool calls: Pearl learns to reason with multimodal tools entirely in the latent space, matching or exceeding SOTA performance without ever explicitly invoking a tool at inference time.