AI & Machine Learning

Introduces a streaming detection head that stops Large Reasoning Models (LRMs) from 'overthinking' redundant steps.

Proposed a test-time scaling paradigm for image restoration that allows compute-to-quality trade-offs during inference.

Releases the hardware design and training environment for MEVIUS2, an open-source, Spot-scale quadruped robot.

Proves that 'topic-matched' contrast pairs are ineffective for extracting refusal directions in LLM abliteration research.

Provides a strictly controlled comparison of autoregressive vs. masked diffusion language models on identical compute budgets.

Ensures safe Vision-Language Model generation without over-refusal by steering activations within the null-space of benign inputs.

Identifies that the direction of log-probability change is more critical than magnitude for improving LLM reasoning via RL.

Integrates LLMs as closed-loop tuning experts for manufacturing robots to achieve 0% failure in complex 3D printing tasks.

Reduces the token count of Stable Diffusion 3.5 by 4x for high-resolution generation with minimal fine-tuning.

Provides causal evidence that LLMs use internal confidence signals to drive behavioral decisions like abstention, rather than just as a side-effect of output generation.

Identifies 'Visual Anchor Collapse' in DPO-aligned VLMs and introduces an asymmetric constraint to prevent models from ignoring visual evidence in favor of language priors.

A predictive scheduling system for multi-agent workflows that optimizes serving across heterogeneous LLM clusters (mixing large and small models).

Introduces 'Noise Titration' to prove that current time-series foundation models often fail at structural inference, behaving instead as 'context parrots' during non-stationary shifts.

Integrates auction bids and monetization logic directly into generative recommender systems (like TIGER) via bid-aware decoding.

MemDLM embeds a simulated denoising process into training to create 'Parametric Memory,' narrowing the train-inference gap for Diffusion Language Models.

An open foundation suite for universal dexterous robot control trained on over 50k trajectories across eight different robotic hand architectures.

Bypasses Reinforcement Learning during the exploration phase by using uncertainty-guided tree search to discover informative data.

Enables high-rank (r=384) DoRA training on single GPUs through factored norms and fused Triton kernels.

Introduces a parallel reasoning mechanism for Vision-Language-Action (VLA) models that eliminates the latency bottleneck of autoregressive Chain-of-Thought.

UNITE enables single-stage joint training of the tokenizer and the diffusion model from scratch, removing the need for frozen VAEs.

A training-free feature caching framework that achieves 2.3x speedup for video world models while maintaining 99.4% quality.

A transformer-based meta-amortized framework that allows simulation-based inference to remain valid across different model structures without retraining.

LassoFlexNet matches or beats leading tree-based models on tabular data while maintaining Lasso-like interpretability through per-feature embeddings and a group Lasso mechanism.

Proves that rotation-invariant algorithms like standard Gradient Descent are fundamentally suboptimal for sparse targets when trained on hard labels.

A grid-free probabilistic framework for nonrigid registration of high-dimensional vector-valued functions on irregular manifolds.

A unified discrete diffusion framework that outperforms autoregressive models on large-scale discrete generation tasks for the first time.

The math we've used for 50 years to figure out how fast the internet should be is actually missing a giant piece of the puzzle.

You can get a whole crowd to agree on something even if everyone only knows what the person right next to them is thinking.

Over 10% of new medical papers are being written by AI now—three years ago, that number was zero.

We can now spot Alzheimer's early by looking at the brain like a building that’s literally buckling under the weight of toxic sludge.

Massive wealth gaps might just be a math problem: if you always pick the better of two random options, inequality is basically guaranteed.

Introduces a statistical alternative to the standard frequency-based BPE tokenization used in nearly all modern LLMs.

Discovers a multiplicative scaling law governing how LLMs revise their beliefs during iterative reasoning (CoT, reflection).

Achieves state-of-the-art LLM distillation using 10-25% of the data required by standard fine-tuning.

Formally proves that a causal Transformer is mathematically equivalent to a stateless Differentiable Neural Computer.

Accelerates MoE inference by speculating future experts to overlap CPU-GPU memory transfers with computation.

A self-improvement framework (MIPO) that improves LLM personalization and reasoning with zero additional data or human labels.

Achieve 97% of Oracle reward performance using only 20% of the training labels for complex LLM reasoning.

The first Joint Embedding Predictive Architecture (JEPA) to train stably end-to-end from raw pixels with massive planning speedups.

Solves the compositional generalization failure of neural networks (0% to 100% accuracy) by embedding algebraic semiring constraints.

A massive controlled study reveals that post-training algorithm rankings (DPO, SimPO, etc.) completely invert as models scale.

DAPA speeds up GELU computation by 16x and reduces hardware DSP utilization by 16x for on-device Transformer deployment.

Spectral Tempering achieves near-oracle embedding compression for dense retrieval without requiring any labeled data or grid searching.

Challenges the 80-year-old assumption that neurons must use weighted summation as their primary aggregation mechanism.

Empirically proves that most Transformer layers are redundant, enabling a 54% training cost reduction through non-uniform budget allocation.

Warm-Start Flow Matching provides a guaranteed speedup for image/text generation by using lightweight models as initial priors.

VAMPO optimizes visual dynamics in video models using policy gradients to fix precision-critical errors in robotic manipulation.

Debunks recent 'evaluation awareness' findings in LLMs by showing that linear probes are actually just tracking formatting artifacts.

Introduces Hyperagents: self-referential systems where the meta-level modification logic is itself an editable program.

Adaptive Layerwise Perturbation (ALP) solves the training-inference mismatch and importance ratio blowup in LLM reinforcement learning.