AI & Machine Learning

Achieves a major breakthrough in dataset distillation, reaching 60% accuracy on ImageNet-1K using only a handful of synthetic images.

Enables 'Elastic Inference' where a single trained model can be converted to multiple lower-precision formats on-the-fly without retraining.

Proposes a parameter-efficient LLM adaptation method that enables rapid specialization on non-stationary streams while preventing catastrophic forgetting.

Replaces manual rubric-tuning for synthetic data with an automated gradient-guided optimization framework based on influence estimation.

Rebuilds the Agent-Computer Interaction (ACI) stack for scientific discovery, solving the fragility of JSON tool-calling and execution sandboxes.

Scales imitation learning data efficiency by generating synthetic 'multi-view' demonstrations from a single expert trajectory.

Introduces SIGN, a framework capable of discovering governing symbolic equations for networked systems with over 100,000 nodes.

Discovers 'Quality Corruption,' an adversarial failure mode where accuracy collapses while detection counts remain stable, proving robustness is substrate-dependent.

Proposes Physical Imitation Learning (PIL) to offload up to 87% of a control policy's mechanical power to passive robotic joints.

OmniVoice is an open-source TTS model scaling to over 600 languages using a novel diffusion language model architecture.

TTA-Vid enables video reasoning models to adapt to new domains at test-time using label-free reinforcement learning on a single sample.

Introduces HiLL, a framework that jointly trains a 'hinter' and 'reasoner' to prevent advantage collapse in reinforcement learning for hard tasks.

Establishes a three-dimensional scaling law for RAG-pretraining, modeling the optimal data budget allocation between model parameters, tokens, and retrieval store size.

CircuitProbe identifies reasoning circuits in Transformers 1000x faster than brute-force methods and predicts the efficacy of layer duplication.

LangMARL introduces agent-level credit assignment and policy gradient evolution directly in the natural language space for multi-agent coordination.

Provides the first controlled study of Silent Data Corruption (SDC) in GPUs and its catastrophic impact on LLM pretraining stability.

Spectral Compact Training (SCT) enables training 70B-parameter architectures on consumer hardware like the Steam Deck (8GB RAM) via permanent SVD factors.

Stochastic Attention achieves a global receptive field in O(log n) layers by using randomized routing inspired by the fruit fly connectome.

ThoughtSteer demonstrates the first successful backdoor attack on continuous latent reasoning models that leave no token-based audit trail.

Mechanistic analysis reveals that LLMs fail at character counting not because they lack the information, but because 'negative circuits' in the final layers actively suppress the correct answer.

This paper achieves O(1) complexity for multimillion-class classification by leveraging predefined vector systems in the latent space.

Routing-Free MoE replaces centralized routing with individual expert-level activation, eliminating the need for Softmax and Top-K load balancing.

Molecular Memory allows MoE systems to recover previously learned domain expertise 9-11x faster by utilizing cost-penalized fitness metrics that preserve dormant experts.

OBD-LLM uses second-order Hessian information to achieve 20-40% better low-rank decomposition accuracy than the current state-of-the-art SVD-LLM.

Policy Improvement Reinforcement Learning (PIRL) shifts the training objective from reward maximization to explicit maximization of policy progress across iterations.

PixelPrune identifies and removes pixel-level redundancy before the Vision Transformer encoder, delivering up to 4.2x inference speedup for high-resolution VLM tasks.

An autonomous research pipeline discovered a lifelong multimodal memory framework by diagnosing and fixing its own architectural bugs and data pipeline issues.

EmbedPart achieves a 100x speedup over Metis for graph partitioning by clustering node embeddings rather than operating on raw graph structures.

A lightweight probing method predicts LLM downstream task performance from internal representations during training, reducing evaluation latency from one hour to three minutes.

Canonical Correlation Analysis (CCA) can reduce image representation dimensionality by 75% while actually improving downstream performance through cross-model agreement.

WARP provides provable, guaranteed repairs for inner layers of Transformers, overcoming the limitation of previous methods restricted to the final layer.

Proposes dense point trajectories as universal 'visual tokens' for behavior that generalize across different species and non-rigid objects.

Releases the GPT-NL Public Corpus, the largest permissively licensed (CC-BY) Dutch-first dataset for LLM pre-training.

Decouples weather forecasting from spatial resolution by using Flow Matching to super-resolve coarse trajectories as a post-processing step.

Solves highly intractable (#P-hard) multi-objective optimization problems with tight approximation guarantees using a novel SAT-oracle approach.

Demonstrates that covert collusion between multi-agent LLM systems can be detected zero-shot using internal model activations.

Achieves 'zero forgetting' in continual learning by stacking frozen domain-specific MoE-LoRA adapters with a meta-router.

First humanoid robot system to achieve consecutive ping-pong strikes using only onboard egocentric vision and whole-body coordination.

Reveals a 'Reasoning Shift' where increased context length silently causes models to skip self-verification and shorten their reasoning traces by up to 50%.

Introduces S0 tuning for hybrid RNN-attention models, outperforming LoRA by 10.8% with zero inference overhead.

Reduces the compute cost of LLM test-time scaling by up to 67% using conformal prediction to calibrate reasoning paths.

Replaces standard relative Softmax attention with 'Multiscreening' to allow absolute query-key relevance, yielding 3.2x faster inference at 100K context.

Simple Self-Distillation (SSD) improves LLM code generation (e.g., Qwen3-30B) by 13% Pass@1 without any external verifiers or teacher models.

Provides causal evidence that reasoning models often decide on an action (like a tool call) before they even start generating their 'Chain-of-Thought'.

Combines the YOCO architecture with recursive computation to scale representational depth without inflating the KV cache.

Solves the long-standing trade-off in low-rank matrix recovery by achieving both optimal sample complexity and fast convergence.

Provides a theoretical explanation for why Transformers often fail compared to linear models in financial time series forecasting.

Enables Gaussian Processes to scale on modern parallel hardware by removing the need for Cholesky decompositions.

Introduces 'deconfounding scores' to enable reliable causal effect estimation even when treatment and control groups have very little overlap.

Delivers a state-of-the-art universal phone recognition model across 100+ languages with full open-source release.