AI & Machine Learning

Fine-tunes Large Vision Language Models for medical tasks using only image-description pairs, bypassing the need for expensive expert-curated instructions.

Introduces Any-Subgroup Equivariant Networks (ASEN), a single model that can adapt to multiple different symmetry groups via input modulation.

ICLAD enables unified, in-context anomaly detection for tabular data across unsupervised, semi-supervised, and one-class regimes without weight updates.

Expands formal reasoning beyond proof construction to the generation and formal verification of counterexamples in Lean 4.

EvidenceRL uses reinforcement learning (GRPO) to explicitly optimize for evidence adherence, reducing hallucinations in high-stakes RAG pipelines.

MoCA3D predicts 3D bounding boxes from monocular images without requiring any camera intrinsics at inference time.

Reveals that complex reasoning strategies like Chain-of-Thought (CoT) and Tree-of-Thought (ToT) provide negligible or even negative gains for text classification tasks.

Formalizes the 'Neural Uncertainty Principle,' linking adversarial vulnerability in vision and hallucinations in LLMs to a shared geometric and information-theoretic origin.

Accelerates diffusion-based image decoders by an order of magnitude using multi-scale sampling and one-step distillation.

CurveStream implements a curvature-aware hierarchical memory to handle streaming video in MLLMs without Out-of-Memory (OOM) errors.

Proves the Key-Value (KV) cache is entirely redundant and can be bit-identically recomputed from the residual stream.

Reduces covariance tracking error by 30x by reformulating the problem as rigid-body motion on Lie groups.

A massive field study (9,000+ users) proves that algorithmic shifts can reduce affective polarization without sacrificing user engagement.

Achieves a 19x reduction in inference cost and 16x in latency for agentic workflows by evolving hybrid LLM-and-code pipelines.

Reduces long-context inference latency by 26.4x using a training-free, structure-aware prompt compression framework.

Boosts open-model agent performance on web navigation tasks from 6.4% to 43%, surpassing proprietary models like GPT-4o.

Proves that intuitive task similarity is a poor predictor of training data value for MLLMs and offers a highly accurate training-free alternative.

Enables zero-shot humanoid robot interaction by generating robot-centric 'dream' videos instead of relying on human-to-robot motion retargeting.

Introduces the first reinforcement learning framework to compress implicit reasoning steps in looped language models.

Replaces fixed context compression ratios with a performance-floor constraint to ensure reliable LLM deployment.

Achieves O(1) time complexity for dense component attribution in SwiGLU Transformers using a single forward-backward pass.

First unified pipeline to reconstruct complete geometry, materials, and lighting from sparse views in under one second.

Introduces the first inherently scalable primitive for radiance fields, allowing real-time Level-of-Detail (LOD) rendering by simply truncating Fourier coefficients.

FIPO overcomes reasoning length stagnation in LLMs by using Future-KL divergence to create dense rewards, extending Chain-of-Thought lengths to over 10,000 tokens.

A training-free method to fix intra-modal misalignment in CLIP by decomposing projectors into an isotropic aligned subspace.

NASimJax provides a 100x throughput increase for autonomous penetration testing simulators by reimplementing the environment in JAX.

SCRL introduces the first negative supervision mechanism for Test-Time Reinforcement Learning, preventing LLMs from reinforcing 'consensus lies'.

SAGE achieves state-of-the-art translation for low-resource languages while reducing training data requirements by 97.1% via RL-guided curation.

Memori reduces agent token costs by 20x by replacing raw conversation history with a persistent layer of semantic triples and summaries.

2K Retrofit enables 2K-resolution inference for any 3D geometric foundation model without modifying or retraining the backbone.

X-World is a controllable, action-conditioned multi-camera world model that simulates realistic future video observations for end-to-end driving.

Breaking the 'capability ceiling' in LLM post-training by replacing full-history dependencies with explicit Markov states.

A k-means variant that is up to 7x faster than FAISS and Scikit-Learn on CPUs and 4x faster than cuVS on GPUs.

Reduces the computational cost of Neural Architecture Search for ensembles from O(M) to O(1).

Enables LLMs to explore beyond their current distribution during RL by treating failed trajectories as hindsight guidance.

Identifies 'critical times' in diffusion generation where targeted guidance pulses significantly improve image control.

Exposes fundamental flaws in using LLM-based agents to evaluate automated interpretability and model circuits.

Replaces unstable free-form recursive LLM code with a typed functional runtime grounded in lambda-calculus.

Derives a variational ELBO for the Joint-Embedding Predictive Architecture (JEPA), unifying it with generative modeling.

Enables zero-shot, directed protein generation by applying a simple scalar bias to stochastic attention samplers.

Demonstrates that LLM reasoning capabilities drop sharply when tasks are framed within multi-turn dialogues vs isolated benchmarks.

A comprehensive end-to-end workflow for humanoid loco-manipulation that standardizes sim-to-real transfer.

Quantifies LLM uncertainty in a single generation pass without auxiliary models or repeated sampling.

Demonstrates that current 'faithfulness' metrics for Chain-of-Thought reasoning are highly subjective and vary wildly depending on the choice of classifier.

Introduces a long-horizon video agent that uses 93% fewer frames than GPT-5/standalone LMMs while achieving higher accuracy.

Provides a robust method for distilling discrete diffusion models that maintains quality and diversity even with very few sampling steps.

Reveals that 'learned priors' in inverse problems often behave as simple lookup tables that memorize training data rather than learning distributions.

Integrates Kolmogorov-Arnold Networks (KANs) into causal generative modeling to produce human-readable symbolic structural equations.

An autonomous AI agent that executes end-to-end theoretical and computational physics research, including hypothesis testing and discovery.

Low-orbit satellites just got scary good—they can pinpoint your location within an inch in basically a heartbeat.