AI & Machine Learning

Reduces high-quality 3D head avatar creation time from over 24 hours to 0.5 seconds per frame.

Reveals that models with identical predictive performance produce fundamentally different feature attributions based solely on their hypothesis class.

Introduces a privacy-preserving ML framework that achieves strong non-invertibility without the utility loss of Differential Privacy or the cost of Homomorphic Encryption.

Fuses categorical sampling into the LM-head matmul to eliminate logit materialization and speed up LLM decoding by up to 19%.

Analyses over 10,000 experiments to prove that LLM agents are capable of genuine architectural discovery rather than just hyperparameter tuning.

Provides empirical evidence that structural sparsity in Vision Transformers does not lead to improved semantic interpretability.

Demonstrates a complete AI-assisted mathematical research loop where a mathematician wrote zero lines of formal code to verify complex physics equilibria.

Integrates LLM agents with the industry-standard Rosetta software to automate physics-based protein design for non-canonical amino acids.

Releases 70B parameter models that operate entirely on bytes, effectively 'liberating' LLMs from static tokenizers.

Derives closed-form power-law scaling for hyperparameters like learning rate and batch size using modern optimization theory rather than expensive empirical sweeps.

Introduces per-token adapter routing, allowing a single sequence to dynamically utilize multiple specialized LoRA experts.

Provides the first formal proof that safety is non-compositional, meaning two individually safe AI agents can become hazardous when combined.

Enables the prediction of an adapter's task, performance, and attributes directly from its LoRA weights without any inference or data access.

Finds that filtering knowledge at 'write-time' (ingestion) maintains 100% RAG accuracy under noise levels where standard 'read-time' filtering completely collapses.

Proposes a protocol that replaces complex multi-agent coding frameworks with a simple, interpretable filesystem structure.

Establishes a duality between sequence-axis attention and depth-wise residual connections, treating layer depth as an ordered variable.

Achieves microsecond-level kinodynamic motion planning for high-DOF robots by using differential flatness to solve boundary value problems analytically.

Introduces ARISE, a hierarchical reinforcement learning framework that allows LLMs to evolve and reuse a tiered library of reasoning skills rather than treating every math problem in isolation.

Challenges the standard use of bilinear/bicubic interpolation for upsampling saliency maps, proving it creates spurious importance regions and proposing a mass-redistribution alternative.

Demonstrates that masked diffusion language models can be 21.8x more compute-efficient than traditional autoregressive models when scaled correctly.

Proposes the Vision-Sound-Language-Action (VSLA) paradigm, enabling robots to respond to real-time environmental acoustics during task execution.

Debunks the widely held 'intra-modal misalignment hypothesis' which claimed CLIP embeddings are inherently poor for image-only tasks.

Introduces Helium, a serving framework that treats agentic workflows as data query plans to optimize redundant LLM calls and KV caches.

Presents ZipCal, a model-agnostic calibration data selection strategy for pruning and quantization that is 240x faster than model-based methods.

Proves that compositional generalization failure in neural networks is an architectural issue and provides a category-theoretic framework to fix it.

Discovers that skipping learning rate decay during pre-training, while appearing worse for pre-train loss, significantly improves the model's adaptability during supervised fine-tuning (SFT).

Proves that noisy/incorrect labels are destructive to Reinforcement Learning with Verifiable Rewards (RLVR), contradicting recent high-profile claims that noise doesn't matter.

Successfully trains a 0.9B parameter pure Spiking Neural Network (SNN) from scratch for language modeling, achieving performance without Transformer distillation.

Formulates Hierarchical Instruction Following as a Constrained Markov Decision Process to ensure LLMs prioritize system prompts over user instructions.

Localizes reinforcement learning updates for code generation by using execution traces to identify the exact point of semantic failure.

Challenges the standard 'pretrain-then-finetune' pipeline by showing that repeating domain-specific data during pretraining is significantly more effective.

A rigorous multi-method audit revealing that frontier LLM performance on MMLU is significantly inflated by data contamination and memorization.

Introduces modular, composable safety alignment via learnable control tokens rather than static parameter-level tuning.

Uses an asymmetric Draft-Verify-Recover pipeline to enable high-quality personalized AI assistants without compromising user privacy.

A self-supervised RLVR method that escapes the 'spurious majority' trap by using a temporary unlearning process for exploration.

Decouples perceptual failures from logical errors in Vision-Language reward models to enable more reliable test-time scaling.

Researchers identified a 'critique vector' in the latent space of Large Reasoning Models that can be steered to improve self-correction and test-time scaling.

Omnilingual MT scales machine translation to over 1,600 languages, an 8x increase in coverage over previous state-of-the-art systems.

This paper demonstrates precise behavioral steering of agentic traits in a 35B parameter MoE model using Sparse Autoencoder (SAE) decoded probe vectors.

FederatedFactory solves the 'extreme non-IID' problem in Federated Learning by federating generative priors instead of model weights.

Laya introduces the first EEG foundation model based on Joint Embedding Predictive Architecture (JEPA), outperforming traditional reconstruction-based models.

Introduces a method to give frozen LLMs persistent memory in their continuous latent space, bypassing the need for text-level RAG or retraining.

IndexRAG shifts cross-document reasoning from inference-time prompting to offline indexing by generating 'bridging facts' at index time.

Provides a theoretical framework for why training AI on what to avoid (negative constraints) is structurally superior and more stable than training on preferences.

VQKV uses Vector Quantization to achieve over 80% KV cache compression with almost zero loss in model performance.

Capability-Guided Compression uses Sparse Autoencoders (SAEs) to prevent 'capability loss' during model pruning and quantization.

A causal analysis reveals that LLMs often ignore their own intermediate reasoning (Chain-of-Thought) when making final decisions.

FEAT is a linear-complexity foundation model designed specifically for extremely large-scale structured (tabular) data.

Kamino is a massively parallel GPU physics solver that natively supports complex kinematic loops and multi-body systems.

Detects and mitigates Vision-Language Model hallucinations at inference time by analyzing visual attention entropy rather than text outputs.