AI & Machine Learning

LongFlow provides an 11x throughput boost for reasoning models by specifically optimizing KV cache for long-output (vs long-input) scenarios.

Manifold-Optimal Guidance reformulates Classifier-Free Guidance (CFG) as a Riemannian control problem, eliminating the artifacts and saturation typical of high guidance scales.

Tiny Aya is a 3.35B parameter multilingual model that achieves state-of-the-art results across 70 languages, challenging the need for massive scale in global AI.

An empirical study reveals that models under 7B parameters have a fundamental utilization bottleneck that prevents them from using retrieved context effectively.

Mobile-GS achieves real-time Gaussian Splatting on mobile devices by replacing the sorting-based alpha-blending bottleneck with depth-aware order-independent rendering.

Expert Threshold Routing (ET) replaces standard top-k token-choice with an independent thresholding mechanism, achieving 1.6x faster training convergence.

RoboClaw introduces 'Entangled Action Pairs' to allow robots to autonomously collect data by learning to reset their own environment.

The discovery of 'Helicoid Dynamics' identifies a critical safety failure where frontier LLMs accurately name their reasoning errors but are structurally unable to stop repeating them.

Achieves 99.5% performance on Needle-In-A-Haystack benchmarks while retaining only 3% of the KV cache budget.

Applying Rotary Positional Embeddings (RoPE) to only 10% of hidden dimensions is sufficient for full model convergence, enabling 10x memory savings in positional caches.

Distills high-fidelity joint audio-visual generation into a real-time streaming model capable of 25 FPS on a single GPU.

Shows that simple sequential fine-tuning with LoRA outperforms complex algorithms for continual reinforcement learning in VLA models.

Proves that policy gradient algorithms naturally collapse entropy and provides a mathematical fix to preserve exploration and diversity.

Achieves hour-scale real-time human animation by solving the unbounded memory growth and inconsistent noise states in autoregressive diffusion.

Introduces the Compression-Consistency Principle, arguing that LLMs prefer truth only when false alternatives are structurally harder to compress.

Replaces unstructured LLM debates with 'Deliberative Collective Intelligence,' producing formal decision packets with minority reports and accountability trails.

Provides a learning-theoretic characterization of model collapse, proving exactly when replaying past outputs destroys model diversity.

Enables agents to autonomously discover the group structure of their environments to learn disentangled representations without human priors.

Unifies leading membership inference attacks into a single framework and uses Bayesian variance inference to enable privacy auditing with 10x less compute.

Automates the entire robotic data generation loop, including a self-resetting mechanism that restores unstructured workspaces without human intervention.

Bridges the gap between parametric CAD and direct B-Rep synthesis using LLMs and primitive grounding.

Eliminates lookahead bias in financial backtesting through a series of yearly-partitioned pretrained LLMs.

Recovers hidden ODE parameters from sparse data with a 487x speedup over gradient-based methods.

Eliminates the 2.5x latency penalty of dynamic adapters in LLMs via pre-gating and fused CUDA kernels.

Enables concurrent perception and reasoning for continuous video streams in Multimodal Large Language Models.

Fits promptable visual segmentation (SAM) into a 1.3M parameter model for real-time in-sensor execution.

First framework for interpreting 4D molecular trajectories into natural language explanations.

Exhaustive circuit mapping of a biological foundation model reveals massive redundancy and annotation bias.

Solves GNN over-squashing by using global effective resistance to identify and rewire structural bottlenecks.

Cross-domain sensor model that handles variable signal lengths and resolutions without retraining.

Achieves high-fidelity one-step (1 NFE) 3D robotic manipulation using training-time drifting fields.

Introduces the first billion-scale SAR vision foundation model and a massive unified benchmark for all-weather geospatial semantic segmentation.

Demonstrates that simply using XML tags during translation outperforms complex pipelines for cross-lingual label projection while actually improving translation quality.

Achieves up to 14.4x higher decoding throughput in long-context LLMs via a training-free framework that reuses sparse memory at semantic boundaries.

Enables multimodal agents to continually improve from experience and skills without any parameter updates through a dual-stream visual grounding framework.

A 3D vision-language pipeline that grounds medical diagnosis in longitudinal brain MRI via regional volumetric assessments to eliminate VLM hallucinations.

Integrates Neural ODEs with NeRFs to enable continuous-time scene dynamics that can extrapolate far beyond the original training sequence.

Proposes a unified image tokenizer that reconciles the conflicting requirements of visual understanding and generation using a residual evolution process.

Identifies and solves the 'information self-locking' failure mode where RL-trained agents stop asking informative questions in active reasoning tasks.

A specialized distributed serving system for 'Any-to-Any' multimodal models that achieves 5.79x lower tail latency via component disaggregation.

Shows that LLM self-correction fails primarily due to 'session context' and can be significantly improved by moving the review to a fresh, independent session.

Automates the generation of GPU-parallelized RL environments from text/code specifications, achieving up to 22,000x speedups for less than $10.

Establishes scaling laws for sampling compute in LLM Reinforcement Learning, providing a playbook for optimal parallel rollout and batch allocation.

Selects high-quality synthetic code data using 'Reverse Mutual Information' to achieve full-dataset performance with 75% less data.

Accelerates sparse attention by 75% by reusing lightning indexer decisions across layers, tackling the hidden bottleneck in production-grade LLMs.

Discovers that task-specific experts are so dense around pretrained weights that random parameter perturbations can compete with complex RL methods like PPO.

Reveals that 'Reasoning LLMs-as-Judges' can lead to policies that generate highly effective adversarial outputs to deceive other judges and inflate benchmarks.

Introduces a feature-matching objective for LLM fine-tuning that targets sequence-level statistics without requiring reward models or ground-truth verifiers.

Integrates Chain-of-Thought reasoning directly into the Diffusion Transformer denoising process to solve complex spatial and logical tasks.

Reduces visual tokens by up to 100x using an autoregressive gazing module, enabling 19x faster 4K/1000-frame video understanding.