Meta AI (FAIR)

On-policy reward modeling with LLM judges not only unlocks significant performance gains on complex mathematical reasoning tasks, but also generalizes to improve performance on simpler numerical and multiple-choice benchmarks.

Ditch the data augmentation and decoders: R2-Dreamer's Barlow Twins-inspired objective delivers faster, more versatile MBRL, especially when spotting the small stuff matters.

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

OmniSONAR halves cross-lingual search error on FLORES and reduces error by 15x on BIBLE, proving that truly universal sentence embeddings across thousands of languages and modalities are now within reach.

Pixel-space diffusion models get a serious boost: V-Co reveals a simple recipe for visual co-denoising that outperforms existing methods on ImageNet-256 with fewer training epochs.

Elastic-Sketch's performance hinges on stream characteristics and eviction thresholds, but this work cracks the code to near-optimal configuration by deriving closed-form expressions for its limiting behavior under stationary random streams.

Forget scaling laws: a specialized 8B parameter translation model can outperform a 70B general-purpose LLM on 1,600 languages.

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.

The optimal spectrogram configuration for audio and speech analysis hinges on a nuanced interplay between front-end feature representation and back-end classifier architecture, varying significantly across tasks.

Current AI's hunger for curated data may be solved by a new architecture inspired by human cognition that flexibly switches between observation, active behavior, and meta-control.

Self-supervised video models can now learn dense features rivaling supervised methods, unlocking a 20-point jump in robot grasping success.

LLMs struggle to generate diverse and specific connections between concepts, even with high token budgets and "thinking" prompts, revealing a gap in creative associative reasoning.

LLM reasoning research is inadvertently paving a dangerous path towards AI situational awareness and strategic deception, demanding a re-evaluation of current safety measures.

Forget imbalanced LoRA usage: ReMix leverages reinforcement learning to route effectively among LoRAs, boosting performance in parameter-efficient fine-tuning.

Even the best open-weight LLMs still fail on nearly two-thirds of questions requiring reasoning over scientific tables, highlighting a persistent "execution bottleneck" in translating strategy to action.

Recursive self-improvement can boost performance by 18% in code and 17% in reasoning, but only if you can keep it from going off the rails – SAHOO provides the guardrails.

Pre-normalization in Transformers is the culprit behind the mysterious link between massive activation outliers and attention sinks, but decoupling them reveals their distinct functions: global parameterization vs. local attention modulation.

Uncover hidden patient subgroups with distinct treatment responses using a new Bayesian clustering approach that goes beyond traditional unsupervised methods.

Datacenter networks are haunted by "ghosts"—topology knowledge failures due to link flaps that occur every 48 seconds at 2025 cluster scale—and existing mitigations are insufficient, but Open Atomic Ethernet offers a potential exorcism.

SSL models can be backdoored with nearly undetectable triggers that still achieve high attack success rates, even against common defenses.

LLMs can ace math problems while reasoning like a drunk toddler, with 82% of correct answers arising from unstable, inconsistent logic.

By disentangling camera-space estimation from world-space refinement via dual diffusion models, DuoMo achieves state-of-the-art human motion reconstruction from noisy video, bypassing the limitations of parametric models.

Stop leaking your face to generative AI apps: PRIVATEEDIT lets you edit images while keeping biometric data on your device.

Safety classifiers for LLMs can catastrophically fail with even minuscule embedding drift, creating dangerous blind spots in deployed safety architectures.

Instruction-following in large reasoning models gets a serious upgrade with RAIN-Merging, a gradient-free technique that merges in instruction-tuned capabilities without wrecking the model's ability to think step-by-step.

AI agents can now learn durable skills instead of constantly "reinventing the wheel," thanks to SkillNet's infrastructure for creating, evaluating, and connecting AI skills at scale.

Unlocking the secrets of viral video ads: a new MLLM framework reveals which initial moments hook viewers and drive conversions.

Randomly throwing distortions at your watermarking model during training? Meta-FC shows meta-learning a better way, boosting robustness by up to 4.71% against combined distortions.

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.

By surgically intervening in MLLM decoding, this work cuts hallucination rates without sacrificing descriptive quality, a feat prior methods struggled to achieve.

A deterministic client selection method leveraging gradient updates can boost federated learning accuracy by nearly 50% in heterogeneous environments.

Escape the bottleneck of translating product intent into ranking system hypotheses: GEARS offers an agentic framework that autonomously discovers and validates superior ranking policies.

Autonomous coding agents derail 30% of the time, but a lightweight intervention system can recover 90% of those misbehaviors with a single nudge.

Achieve zero-collision embedding tables in production recommenders without sacrificing training speed, unlocking better personalization via fresher and higher-quality item embeddings.

Forget quadratic complexity: ULTRA-HSTU achieves 21x faster inference and 4-8% better engagement in large-scale recommendation by co-designing input sequences, sparse attention, and model topology.

Despite growing interest, queer NLP research remains largely reactive, highlighting biases instead of building proactive solutions, leaving significant opportunities for stakeholder-driven and intersectional approaches.

Ditch ANN search altogether: MFLI learns a hierarchical index alongside item embeddings, boosting recall by up to 11.8% and cold-content delivery by 57.29% in large-scale recommender systems.

Unlock superhuman visual reasoning in multimodal models by simply giving them the ability to think step-by-step at test time.

Object-level masking in world models unlocks a 20% boost in counterfactual reasoning and drastically reduces planning costs, hinting at a path toward more efficient and robust AI agents.

Forget noisy user data: synthetic data unlocks predictable scaling laws for LLMs in recommendation, boosting recall by 130%.

Forget huge models: parameter-efficient fine-tuning turns tiny language models into code-generating powerhouses that outperform larger, untuned counterparts.

Achieve state-of-the-art UAV detection by swapping transformers for Mamba, yielding a faster and more accurate multimodal detector.

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.

LLM safety guardrails are far less robust than benchmarks suggest, with accuracy dropping by as much as 57% on novel adversarial attacks, and some even generating harmful content in a "helpful mode" jailbreak.

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.

Forget scaling laws: this work shows you can get SOTA reasoning from sub-billion parameter models with *less* data, if you're smart about curation and resampling.

Dramatically slash the carbon footprint of AI training without sacrificing performance by co-designing hardware and software for modern GPUs.

Achieve up to 39.6% FLOP reduction in LLM inference without retraining or architectural changes using QuickSilver's dynamic token-level optimizations.

Ditch the pre-trained models: PAST directly learns speech tokens from phonetic data, outperforming existing methods in representation and reconstruction.

LLMs can generate plain language summaries of scientific research that are as good as human-written ones, but easier to read.

Edit the bassline, drums, or other instruments of any song with this new open-source multi-stem music generation model.