AI & Machine Learning

Quantifies near-verbatim data extraction risk in LLMs at 1/5000th the computational cost of standard Monte Carlo methods.

Enables graph-based retrieval and reranking for RAG without the maintenance overhead of a knowledge graph.

Reduces visual tokens in robot policies by 78% by using inter-layer rank consistency instead of simple attention magnitude.

This paper demonstrates that the order of training examples alone can encode information not present in any individual example, allowing models to bypass established sample complexity bounds.

A systematic study reveals that grokking is not an architectural property of Transformers but an interaction between weight decay and optimization stability.

The 'Reasoning Contamination Effect' shows that Chain-of-Thought (CoT) reasoning actually disrupts a model's internal confidence signal, leading to poorer calibration.

Large Language Models process instructions as social acts rather than technical specifications, making 'imperative mood' prompts behave inconsistently across different languages.

GeoNDC introduces a queryable neural data cube that compresses 20 years of planetary satellite data by 95x while allowing on-demand continuous-time reconstruction.

Sparton is a specialized Triton kernel that solves the massive memory bottleneck of Learned Sparse Retrieval (LSR) models like Splade.

Intern-S1-Pro is the first trillion-parameter scientific multimodal foundation model, outperforming proprietary models on specialized scientific reasoning.

AirVLA successfully transfers manipulation-trained Vision-Language-Action (VLA) models to underactuated aerial robots using a payload-aware guidance mechanism.

R1Sim applies the 'Reasoning-RL' paradigm (popularized by DeepSeek-R1) to traffic simulation, achieving superior safety and diversity in multi-agent behaviors.

SIGMA resolves 'trajectory divergence' in molecular string generation by enforcing geometric symmetry recognition through contrastive learning.

A fully differentiable agent-based traffic simulator enables calibration and control of million-vehicle networks 173x faster than real-time.

GIFT is a training-free frame selection framework that uses 'Directed Diversity' to boost Video-LLM performance by up to 12.5%.

Z-Erase introduces the first concept erasure method for single-stream diffusion transformers, preventing generation collapse in new unified architectures.

This paper demonstrates that Sparse Autoencoder (SAE) features in multimodal models are not modular, challenging the core assumption of intervention-based steering.

Pixelis shifts VLM reasoning from static description to a 'reasoning in pixels' agentic paradigm that learns via an executable tool grammar.

The AE4E paradigm proposes a 'Social Contract' for multi-agent economies, replacing individual model alignment with an institutional 'Separation of Power'.

MSRL scales multimodal reward modeling by transferring reasoning capabilities from text to vision-language tasks without requiring new multimodal preference data.

SEVerA enables the synthesis of self-evolving agents with formal guarantees by combining LLM planning with first-order logic rejection samplers.

Using Signal Detection Theory, this work proves that LLM calibration and 'metacognitive efficiency' (knowing what you know) are distinct, dissociable capacities.

Photon enables efficient 3D medical volume understanding through adaptive token scheduling and a novel 'gradient restoration' backpropagation rule.

Vision Hopfield Memory Networks (V-HMN) present a brain-inspired alternative to Transformers and Mamba using hierarchical associative memory mechanisms.

Trace2Skill distills lessons from across a 'parallel fleet' of execution trajectories into a unified, conflict-free skill directory for LLM agents.

Pruning low-utility prompts before RL rollouts allows for 10x more efficient training of large reasoning models.

Safety alignment does not have to be a 'tax' on performance; it can actually improve mathematical reasoning accuracy.

Enable long video generation from short-video diffusion models without any additional training or fine-tuning.

Training-free 6D pose estimation for unseen surgical instruments using only a CAD model as prior knowledge.

Offline Decision Transformers can now synthesize strategies that surpass the classical heuristics they were trained on for the Traveling Salesman Problem.

Simple image sharpening serves as a surrogate-free, zero-cost preemptive defense against adversarial attacks.

Representing GPS trajectories as hyperspectral images enables multi-month dense anomaly detection that was previously computationally intractable.

A foundation model for gait transforms 3D skeletal motion into a systemic biosignal for multi-system health monitoring.

A new tokenization architecture reduces the 'Token Tax' for complex non-Latin scripts by over 60%.

Sparse Autoencoder analysis reveals that weight pruning counter-intuitively preserves rare features better than frequent ones.

LLMs can be fine-tuned to act as their own 'Z-token' compressors, achieving 18x text reduction without losing reconstruction fidelity.

GlowQ introduces group-shared low-rank approximations to speed up quantized LLM inference by up to 37%.

Defines 'Reasoning Safety' as a new security dimension and introduces a real-time monitor to detect logic-chain hijackings.

Cross-model disagreement (CMP/CME) provides a highly effective, label-free signal for detecting confident hallucinations.

Introduces a training-free pipeline for pixel-level video anomaly detection that achieves a 5x improvement in object-level accuracy.

A model-agnostic framework to extract the model-implied causal structure from any trained temporal predictor.

Reduces LLM inference energy by 40% (and up to 81%) using a distillation-based router to skip unnecessary reasoning steps.

Detects when object detectors fail to see safety-critical objects by measuring semantic misalignment with foundation model embeddings.

Challenges the 'Golden Data' requirement for video generation by showing that imbalanced data can outperform high-quality data through timestep-aware training.

Unlocks full-body musculoskeletal humanoid training by achieving order-of-magnitude speedups via massively parallel GPU simulation.

Fixes the inherent instability of on-policy distillation in LLMs using local support matching and top-p rollout sampling.

Achieves 45% performance gains in robotics using 5-10x fewer real-world demonstrations through high-dimensional factorization.

Enables LMMs to 'think' using compact latent visual representations rather than verbalizing everything into text.

Translates a single natural language sentence into a validated, hardware-specific computational imaging system design.

Achieves up to 4.7x speedup for Diffusion LLMs using a training-free self-speculative decoding framework.